Deep learning has become a transformational technology in the field of medical image analysis, offering powerful tools for diagnosis of disease, segmentation, and planning treatment. This paper outlines a comprehensive study on the use of deep learning on medical images, with a specific focus on a new preprocessing machine learning algorithm that enhances overall model performance. We introduce the Adaptive Multi-Scale Medical Image Preprocessing Algorithm (AMSMIPA), a robust pipeline that adapts denoising, multi-modal contrast-enhancement, and edge-preserving smoothing to standardize and improve the quality of medical images from multiple modalities. We evaluate our method's effectiveness on two large-scale datasets: the Stanford AIMI CheXpert Plus dataset and the MedSegBench benchmark. We implement and train five state-of-the-art deep learning architectures: our own custom convolutional neural network (CNN), ResNet, DenseNet, a Vision Transformer (ViT), and U-Net. Our results demonstrate our preprocessing algorithm greatly enhances model performance to be significantly higher with all architectures. When used in combination with our AMSMIPA preprocessing, the Vision Transformer (ViT) model achieved the highest performance at 92.1% accuracy, a macro F1-score of 0.921, and a ROC-AUC of 0.994. This study provides a thorough analysis of deep learning techniques for medical image processing and highlights the critical role of advanced preprocessing in achieving state-of-the-art results. All code, datasets, and results are made publicly available to encourage further research and development in this field. .............................. Keywords:.......................... Medical image analysis, Deep learning, Preprocessing algorithms, Vision Transformer (ViT), U-Net